A model for perceptual category information

Abstract

The purpose of this investigation was the development and preliminary testing of a model for perceptual category formation in the absence of information as to the correctness of category assignment. The development was based on the assumption that potentially useful categories are indicated in the visible environment by association with a large number of covarying perceptual characteristics. A theoretical machine was proposed to describe a process by which category formation consistent with this assumption could occur. The theoretical machine was developed under restrictions intended (1) to make the machine compatible with a statistical model for learning, and (2) to facilitate the exploration of relationships between the requirements of the model and neurophysiological theory. Investigation of the basic assumption entailed the preparation of a computer program which could generate patterns having the characteristics specified by the assumption. The theoretical machine was then simulated in the form of a computer program and presented with the patterns; the machine demonstrated that it could indeed form categories consistent with the basic assumption. The patterns were also presented to human subjects with the request that they assign the patterns to one of two categories. Some, but not all, of the human subjects showed a significant tendency to categorize in a fashion consistent with the assumption. An effort was made to find parameters for the model which would cause it to produce categorizing performance similar in detail to that of the human subjects. Some gross correspondences between the two performances were achieved, but the model exhibited an irremediable tendency to achieve consistent categorization more rapidly than humans did. Consideration of the requirements imposed on the human and the model suggested that the model had a less difficult task than did the humans and might more nearly approximate the human performance if the model were modified to allow it to undertake a task more nearly comparable to that of the human subjects.